Soil recovery and occasional tillage in regenerative agricultural systems of the Brazilian Cerrado: carbon dynamics, proximal sensing and machine learning

Abstract

The conversion of native vegetation to intensive agricultural systems in the Brazilian Cerrado has resulted in substantial losses of soil organic carbon, compromising physical properties essential to agroecosystem functioning and generating a carbon debt that poses a challenge to the sustainability of tropical agriculture. Although regenerative agriculture is recognized as a promising strategy for restoring soil health, knowledge gaps remain regarding the recovery of organic carbon fractions following occasional mechanical disturbances and the methods available to monitor these changes. Therefore, this thesis aimed to understand the recovery of soil organic carbon and soil quality in regenerative soybean–cotton systems in the Brazilian Cerrado and to evaluate the potential of integrating proximal sensors, including portable X-ray fluorescence (pXRF), visible and near-infrared diffuse reflectance spectroscopy (Vis-NIR), and Fourier-transform infrared spectroscopy (FTIR), with machine-learning algorithms to predict carbon stocks and soil physical and chemical properties. The study was conducted on Oxisols in the Parecis region of Mato Grosso, using areas of native Cerrado vegetation and agricultural fields under regenerative no-tillage with one, six, and ten years of recovery following a single strategic subsoiling operation performed to alleviate soil compaction. Soil samples collected to a depth of 1.0 m were used to determine total soil organic carbon stocks, particulate organic carbon (POC), mineral-associated organic carbon (MAOC), total nitrogen, and soil physical and chemical properties, in addition to acquiring elemental and spectral information from the sensors. The results showed that carbon recovery differed among carbon fractions and throughout the soil profile. MAOC showed progressive recovery, particularly below 20 cm, approaching or exceeding the values observed under native vegetation in parts of the soil profile. In contrast, POC remained the main component of the carbon debt after ten years of recovery, indicating that its replenishment depends on sustained increases in biomass inputs and the rebuilding of soil structure. Gradual improvements in soil aggregation, porosity, and water retention were also observed, indicating that the recovery of soil functionality is associated with changes in soil structure and carbon stabilization processes. The predictive models showed high performance in estimating carbon stocks, functional organic matter fractions, nitrogen, and several soil physical and chemical properties, demonstrating that the integration of pXRF, FTIR, and Vis-NIR improves predictive capability compared with the use of individual sensors. The results demonstrate that regenerative agriculture promotes the gradual restoration of soil functionality in Cerrado Oxisols and that the combination of proximal sensing and machine learning provides a robust approach for monitoring soil health. By integrating carbon stabilization processes with advanced diagnostic technologies, this thesis advances the understanding of tropical soil regeneration and provides a scientific basis for low-carbon agriculture programs, monitoring, reporting and verification (MRV), environmental certification, and sustainability assessment of regenerative agricultural systems.